Style translation of simulated ECGs to realistic ECGs with generative adversarial networks for data-scarce clinical settings: challenges and opportunities

AI-based predictive models on electrocardiograms (ECGs) for rare diseases underperform because labelled data are scarce. Simulated ECGs from electrophysiological models embed disease-specific features, yet clinicians still recognise their synthetic style, and naïve inclusion of them in training sets degrades performance.

We frame augmentation with simulated ECGs as a style-transfer task—content is correct, style is artificial. We test whether adversarial style transfer narrows this domain gap and outline why diffusion models may overcome the shortcomings of adversarial ones.

CycleGAN translated simulated signals into a more realistic style while retaining diagnostic information. Using 987 PTB-XL ECGs (56 right bundle-branch block, RBBB) plus 25 RBBB simulations from MedalCareXL, we trained classifiers on (i) real data only, (ii) real + raw simulations, and (iii) real + translated simulations. We also analysed convergence limitations of adversarial versus diffusion approaches.

Raw simulations impaired performance (area under the curve, AUC 0.93; average precision, AP 0.68). CycleGAN-translated data restored performance to the real-only baseline (AUC 0.95 vs 0.96, AP 0.81 vs 0.79) but remained susceptible to unstable training.

Style transfer mitigates the harm of simulated ECGs, yet GAN instability restricts clinical reliability. Diffusion models, which provide objective convergence diagnostics, may offer a safer route for ECG style transfer in data-scarce medical AI.